Improving Linux Block I/O for Enterprise Workloads

The block I/O subsystem of the Linux kernel is one of the critical components affecting the performance of server workloads. Servers typ-ically scale their I/O bandwidth by increasing the number of attached disks and controllers. Hence, the scalability of the block I/O layer is also an important concern. In this paper, we examine the performance of the 2.4 Linux kernel’s block I/O subsystem on enterprise workloads. We identify some of the major bottlenecks in the block layer and propose kernel modifications to alleviate these problems in the context of the 2.4 kernel. The performance impact of the proposed patches is shown using a decision-support workload, a microbenchmark, and profiling tools. We also examine the newly rewritten block layer of the 2.5 kernel to see if it addresses the performance bottlenecks discovered earlier.